CDF Run II Offline Performance Test

Goal

We will describe the performances tests executed on a Linux machine (pcdfsrv.pd.infn.it) dedicated for CDF Run II code development and analysis located in Padova. The performance tests will focus on:

compiling and linking time of executables.
executing times for a typical physics analysis job.

Particular attention has been paid also to run the tests in heavy load conditions.

Hardware and Software Description

The pcdfsrv.pd.infn.it machine is a Linux quadripocessor Intel box.

The version of the CDF RunII offline is 3.16.0 using the shared libraries and all the tests will be performed using NNbTag (my own version of ExoticMods). This module builds a job that performs the following features (i.e. a job for a physical analysis):

Jet clustering (Cone Algorithm).

Tracking in the COT.

Tracking in Silicon (OutsideIn).

SecVtx b-tag.

Jet Probability.

Calculates some jet-related useful variables.

To load the NNbTag package do the following steps:

setup cdfsoft2 3.16.0
setenv USESHLIBS 1
newrel -t 3.16.0 test
cd test
addpkg ExoticMods
download nnbtag.tar
gunzip nnbtag.tar.gz
tar -xvf nnbtag.tar

That will generate a package whose name is ExoticMods but it will be completly different from the one that you can upload from the CVS. The NNbTag package will be added to the standard release as soon as I have time to do it! So in the following steps I'll refer to ExoticMods rather than NNbTag. It is NOT a mistake.

Test A: Compiling and Linking

We have measured:

the elapsed "ClockWall" time ( %E in time command of csh) for compiling the whole ExoticMods package to generate the ExoticModslib libraries (gmake ExoticMods.nobin).
the "Clock Wall" time to link and build the executable (gmake ExoticMods.tbin).

It might be interesting to understand wether one can use the CDFII offline under AFS to compile, link and build executables. So we have repeated the previous tests using the CDF II offline on AFS located in two different geographical locations: Trieste and CNAF (Bologna). The last setup should be faster since the CNAF connection is linked with a wider band.

For comparison the same tests has been performed on the fcdflnx1 box at FNAL. The results are summarized in the following table.

	pcdfsrvr local	pcdfsrvr AFS TS	pcdfsrvr AFS BO	fcdflnx1
gmake ExoticMods.nobin	2:56	4:51	5:03	2:43
gmake ExoticMods.bin	3:01	51:42	37:42	2:55

Summary:

Compiling and linking time are approximately the same using fcdflnx1 or pcdfsrvr compiled with local libraries.
The time for linking using remote libraries is approximately ten times more using local libraries
An improvement using the remote libraries can be obtained increasing the bandwith.

Test B: Generation of events

With the standard cdfSim executable we have measured the time to generate 50 ttbar events with IsaJet (ttbar.tar). To check the performances in high load conditions we have measured the performances while 1,2,4 and 8 processes were running at the same time. The results are summarized in the pictures below.

Summary:
During the generation of ttbar events the CPU are well occupied by the processes. In particular the "Wall Clock" time doubles when the number of processes is twice the nuber of precessors. That means that no time is loss shuffling around the processes.

Test C: Reconstruction of Events

On 50 z->bbbar events we have run the NNbTag job described before. Also in this case we have checked the behaviour of the system in a heavy load situation. Also the NNbTag job is a very complete job that performs jet reconstruction, silicon clustering, tracking and search for secondary vertex. As in the previous test we have run 1,2,4, and 8 process in the same time. The results are summarized below.

Summary:
Also in this case, where more memory is used, the "Wall Clock" seems to scale well when the number of processes is twice the number of processor (203.75s for 4 processes against 414.56s for 8 processes).

TestD: I/O Performances

After the reconstruction is executed the CdfTrackCollection are stored using STNtuple package. The time to write them is measured also in heavy load condition. Also the time to read them using a root script is evaluated.

TODO

Conclusions:

TODO

Antonio Sidoti me fecit
Last Update: 03 July 2001

Back to the RunII Computing and Software Page